Spaces:
Sleeping
Sleeping
Update model.py
Browse files
model.py
CHANGED
@@ -1,8 +1,26 @@
|
|
1 |
-
import math
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
-
|
5 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
class RMSNorm(torch.nn.Module):
|
8 |
def __init__(self, dim: int, eps: float = 1e-6):
|
@@ -17,13 +35,12 @@ class RMSNorm(torch.nn.Module):
|
|
17 |
output = self._norm(x.float()).type_as(x)
|
18 |
return output * self.weight
|
19 |
|
20 |
-
|
21 |
class Attention(nn.Module):
|
22 |
"""
|
23 |
Multi-head Self-Attention with RoPE
|
24 |
"""
|
25 |
|
26 |
-
def __init__(self, num_heads, head_size, num_embed
|
27 |
super().__init__()
|
28 |
self.num_heads = num_heads
|
29 |
self.head_size = head_size
|
@@ -32,27 +49,8 @@ class Attention(nn.Module):
|
|
32 |
self.wk = nn.Linear(num_embed, num_heads * head_size, bias = False)
|
33 |
self.wv = nn.Linear(num_embed, num_heads * head_size, bias = False)
|
34 |
self.wo = nn.Linear(num_heads * head_size, num_embed, bias = False)
|
35 |
-
|
36 |
-
|
37 |
-
self.register_buffer('inv_freq', inv_freq)
|
38 |
-
|
39 |
-
self.dropout = nn.Dropout(dropout)
|
40 |
-
|
41 |
-
def reshape_for_broadcast(self, freq_cis, x):
|
42 |
-
ndim = x.ndim
|
43 |
-
shape = [1] * (ndim - 2) + list(freq_cis.shape)
|
44 |
-
return freq_cis.view(*shape)
|
45 |
-
|
46 |
-
def apply_rope(self, x, position, freq):
|
47 |
-
t = torch.arange(position, device=freq.device, dtype=torch.float32)
|
48 |
-
freq = torch.outer(t, freq)
|
49 |
-
freq_cis = torch.polar(torch.ones_like(freq), freq)
|
50 |
-
x_ = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
|
51 |
-
freq_cis = self.reshape_for_broadcast(freq_cis, x)
|
52 |
-
x_out = torch.view_as_real(x_ * freq_cis).flatten(3)
|
53 |
-
return x_out.type_as(x)
|
54 |
-
|
55 |
-
def forward(self, x):
|
56 |
B, T, C = x.shape
|
57 |
|
58 |
mask = torch.triu(torch.full((T, T), float("-inf"), device=x.device), diagonal=1)
|
@@ -67,91 +65,62 @@ class Attention(nn.Module):
|
|
67 |
xk = xk.transpose(1, 2)
|
68 |
xv = xv.transpose(1, 2)
|
69 |
|
70 |
-
xq =
|
71 |
-
xk =
|
72 |
|
73 |
attn_weights = torch.matmul(xq, xk.transpose(2, 3)) / math.sqrt(self.head_size)
|
74 |
attn_weights += mask
|
75 |
attn_weights = F.softmax(attn_weights.float(), dim=-1).type_as(xq)
|
76 |
output = torch.matmul(attn_weights, xv)
|
77 |
output = output.transpose(1, 2).contiguous().view(B, T, C)
|
78 |
-
return self.
|
79 |
-
|
80 |
|
81 |
class MLP(nn.Module):
|
82 |
-
"""
|
83 |
-
Implementation of a Multi-Layer Perceptron (MLP) sub-module.
|
84 |
-
|
85 |
-
This module is a simple feed-forward network with two hidden layers
|
86 |
-
used in various Transformer components like the Mixture of Experts layer.
|
87 |
-
"""
|
88 |
-
|
89 |
def __init__(self, num_embed, dropout):
|
90 |
-
"""
|
91 |
-
Constructor for the MLP.
|
92 |
-
|
93 |
-
Args:
|
94 |
-
num_embed (int): The number of embedding dimensions.
|
95 |
-
"""
|
96 |
-
|
97 |
super().__init__()
|
98 |
-
|
99 |
|
100 |
-
|
101 |
-
self.w1 = nn.Linear(num_embed, hidden, bias=False)
|
102 |
-
self.w2 = nn.Linear(hidden, num_embed, bias=False)
|
103 |
|
|
|
|
|
104 |
self.dropout = nn.Dropout(dropout)
|
105 |
|
106 |
def forward(self, x):
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
x
|
112 |
-
|
113 |
-
Returns:
|
114 |
-
torch.Tensor: Output tensor after passing through the MLP (shape: batch_size, seq_len, num_embed).
|
115 |
-
"""
|
116 |
-
return self.dropout(self.w2(F.silu(self.w1(x))))
|
117 |
-
|
118 |
class TransformerBlock(nn.Module):
|
119 |
"""
|
120 |
This calss will group together MultiHead Attention and
|
121 |
-
|
122 |
"""
|
123 |
|
124 |
-
def __init__(self, num_heads,
|
125 |
super().__init__()
|
126 |
-
|
127 |
-
self.
|
|
|
|
|
128 |
num_heads=num_heads,
|
129 |
head_size=head_size,
|
130 |
-
num_embed=num_embed
|
131 |
-
dropout=dropout
|
132 |
)
|
133 |
-
|
134 |
-
self.mlp = MLP(num_embed = num_embed, dropout = dropout)
|
135 |
-
|
136 |
# add the layer normalization
|
137 |
-
self.
|
138 |
-
self.
|
139 |
-
|
140 |
-
def forward(self, x):
|
141 |
-
""
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
Returns:
|
149 |
-
torch.Tensor: A tensor of shape (batch_size, sequence_length, embedding_dim).
|
150 |
-
"""
|
151 |
-
#print(x.shape)
|
152 |
-
x = x + self.mha(self.norm1(x))
|
153 |
-
x = x + self.mlp(self.norm2(x))
|
154 |
-
|
155 |
return x
|
156 |
|
157 |
|
@@ -161,82 +130,80 @@ class Transformer(nn.Module):
|
|
161 |
# a simple lookup table that stores embeddings of a fixed dictionary and size
|
162 |
# each token directly reads off the logits for the next token from a lookup table
|
163 |
# see more: https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html
|
164 |
-
self.model_type = 'Prome'
|
165 |
self.vocab_size = kwargs.get("vocab_size", 100)
|
166 |
self.num_embed = kwargs.get("num_embed", 32)
|
167 |
-
self.block_size = kwargs.get("block_size", 8)
|
168 |
self.num_heads = kwargs.get("num_heads", 4)
|
169 |
-
self.head_size = kwargs.get("head_size", 128)
|
170 |
self.num_layers = kwargs.get("num_layers", 4)
|
|
|
171 |
self.dropout = kwargs.get("dropout", 0.2)
|
172 |
-
self.max_seq_len = kwargs.get("max_sqe_len", 1024)
|
173 |
# each token reads the logits for the next token from a lookup table
|
174 |
self.token_embedding_table = nn.Embedding(self.vocab_size, self.num_embed)
|
175 |
# each position from 0 to block_size-1 will get its embedding
|
176 |
-
#self.position_embedding_table = nn.Embedding(self.
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
for _ in range(self.num_layers)
|
187 |
-
]
|
188 |
-
)
|
189 |
-
|
190 |
self.lm_head = nn.Linear(self.num_embed, self.vocab_size)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
|
192 |
def forward(self, idx, targets=None):
|
193 |
B, T = idx.shape
|
194 |
# idx and targets are (B,T) tensor of integers
|
195 |
# the token_emb is (B, T, C), C = NUM_EMBED
|
196 |
x = self.token_embedding_table(idx)
|
197 |
-
# (T, C)
|
198 |
-
#posit_emb = self.position_embedding_table(torch.arange(T, device=DEVICE))
|
199 |
-
|
200 |
-
#x = token_emb + posit_emb
|
201 |
|
202 |
-
|
|
|
|
|
|
|
203 |
|
|
|
|
|
204 |
# (B, T, vocab_size)
|
205 |
logits = self.lm_head(x)
|
206 |
-
|
207 |
-
# Compute the loss
|
208 |
if targets != None:
|
209 |
# cross_entropy accepts inputs in a (batch_size, num_classes)
|
210 |
# so we need to reformat our logits dimensions to
|
211 |
# (batch_size * time, dim_vocabulary), time = block_size
|
212 |
-
#logits = logits.to(dtype=torch.float32)
|
213 |
-
|
214 |
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
|
215 |
else:
|
216 |
loss = None
|
217 |
-
|
218 |
return logits, loss
|
219 |
|
220 |
-
def generate(self, idx: torch.Tensor, max_new_tokens: int, temperature: float = 0.
|
221 |
for _ in range(max_new_tokens):
|
222 |
idx_crop = idx[:, -self.max_seq_len:]
|
223 |
|
|
|
224 |
logits, loss = self.forward(idx_crop)
|
225 |
logits = logits[:, -1, :]
|
226 |
|
227 |
if temperature > 0:
|
228 |
-
probs = F.softmax(logits / temperature, dim=-1)
|
229 |
idx_next = self.sample_top_p(probs, top_p)
|
230 |
else:
|
231 |
probs = F.softmax(logits, dim=-1)
|
232 |
idx_next = torch.multinomial(probs, num_samples=1)
|
233 |
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
|
234 |
-
return idx
|
235 |
|
236 |
def sample_top_p(self, probs: torch.Tensor, top_p: float) -> torch.Tensor:
|
237 |
sorted_probs, sorted_indices = torch.sort(probs, descending=True, dim=-1)
|
238 |
cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
|
239 |
-
|
240 |
# Create a mask for top-p filtering
|
241 |
top_p_mask = cumulative_probs <= top_p
|
242 |
top_p_mask[..., 1:] = top_p_mask[..., :-1].clone()
|
|
|
1 |
+
import math
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
+
import torch.nn.functional as F
|
5 |
+
|
6 |
+
def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
|
7 |
+
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
|
8 |
+
t = torch.arange(end, device=freqs.device)
|
9 |
+
freqs = torch.outer(t, freqs)
|
10 |
+
freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64
|
11 |
+
return freqs_cis
|
12 |
+
|
13 |
+
def reshape_for_broadcast(freqs_cis, x):
|
14 |
+
batch_size, num_heads, seq_len, head_size = x.shape
|
15 |
+
freqs_cis = freqs_cis[:seq_len]
|
16 |
+
shape = [1, 1, seq_len, head_size // 2]
|
17 |
+
return freqs_cis.view(*shape)
|
18 |
+
|
19 |
+
def apply_rope(x, position, freqs_cis):
|
20 |
+
x_ = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
|
21 |
+
freqs_cis = reshape_for_broadcast(freqs_cis, x)
|
22 |
+
x_out = torch.view_as_real(x_ * freqs_cis).flatten(3)
|
23 |
+
return x_out.type_as(x)
|
24 |
|
25 |
class RMSNorm(torch.nn.Module):
|
26 |
def __init__(self, dim: int, eps: float = 1e-6):
|
|
|
35 |
output = self._norm(x.float()).type_as(x)
|
36 |
return output * self.weight
|
37 |
|
|
|
38 |
class Attention(nn.Module):
|
39 |
"""
|
40 |
Multi-head Self-Attention with RoPE
|
41 |
"""
|
42 |
|
43 |
+
def __init__(self, num_heads, head_size, num_embed):
|
44 |
super().__init__()
|
45 |
self.num_heads = num_heads
|
46 |
self.head_size = head_size
|
|
|
49 |
self.wk = nn.Linear(num_embed, num_heads * head_size, bias = False)
|
50 |
self.wv = nn.Linear(num_embed, num_heads * head_size, bias = False)
|
51 |
self.wo = nn.Linear(num_heads * head_size, num_embed, bias = False)
|
52 |
+
|
53 |
+
def forward(self, x, freqs_cis):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
B, T, C = x.shape
|
55 |
|
56 |
mask = torch.triu(torch.full((T, T), float("-inf"), device=x.device), diagonal=1)
|
|
|
65 |
xk = xk.transpose(1, 2)
|
66 |
xv = xv.transpose(1, 2)
|
67 |
|
68 |
+
xq = apply_rope(xq, T, freqs_cis)
|
69 |
+
xk = apply_rope(xk, T, freqs_cis)
|
70 |
|
71 |
attn_weights = torch.matmul(xq, xk.transpose(2, 3)) / math.sqrt(self.head_size)
|
72 |
attn_weights += mask
|
73 |
attn_weights = F.softmax(attn_weights.float(), dim=-1).type_as(xq)
|
74 |
output = torch.matmul(attn_weights, xv)
|
75 |
output = output.transpose(1, 2).contiguous().view(B, T, C)
|
76 |
+
return self.wo(output)
|
|
|
77 |
|
78 |
class MLP(nn.Module):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
def __init__(self, num_embed, dropout):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
super().__init__()
|
81 |
+
self.num_embed = num_embed
|
82 |
|
83 |
+
hidden_dim = 3 * int(num_embed * 2 / 3)
|
|
|
|
|
84 |
|
85 |
+
self.linear1 = nn.Linear(num_embed, hidden_dim)
|
86 |
+
self.linear2 = nn.Linear(hidden_dim, num_embed)
|
87 |
self.dropout = nn.Dropout(dropout)
|
88 |
|
89 |
def forward(self, x):
|
90 |
+
x = self.linear1(x)
|
91 |
+
x = F.silu(x)
|
92 |
+
x = self.linear2(x)
|
93 |
+
x = self.dropout(x)
|
94 |
+
return x
|
95 |
+
|
|
|
|
|
|
|
|
|
|
|
96 |
class TransformerBlock(nn.Module):
|
97 |
"""
|
98 |
This calss will group together MultiHead Attention and
|
99 |
+
FeedForward NN, so that we can copy it in Transformer
|
100 |
"""
|
101 |
|
102 |
+
def __init__(self, num_heads, num_embed, dropout):
|
103 |
super().__init__()
|
104 |
+
self.num_heads = num_heads
|
105 |
+
self.num_embed = num_embed
|
106 |
+
head_size = num_embed // num_heads
|
107 |
+
self.sa = Attention(
|
108 |
num_heads=num_heads,
|
109 |
head_size=head_size,
|
110 |
+
num_embed=num_embed
|
|
|
111 |
)
|
112 |
+
self.ffwd = MLP(num_embed=num_embed, dropout=dropout)
|
|
|
|
|
113 |
# add the layer normalization
|
114 |
+
self.ln1 = RMSNorm(num_embed)
|
115 |
+
self.ln2 = RMSNorm(num_embed)
|
116 |
+
|
117 |
+
def forward(self, x, freqs_cis):
|
118 |
+
# "x +" is the skip (or residual) connection
|
119 |
+
# it helps with optimization
|
120 |
+
# also we apply layer normalization before self-attention
|
121 |
+
# and feed-forward (a reshufle from original paper)
|
122 |
+
x = x + self.sa(self.ln1(x), freqs_cis)
|
123 |
+
x = x + self.ffwd(self.ln2(x))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
return x
|
125 |
|
126 |
|
|
|
130 |
# a simple lookup table that stores embeddings of a fixed dictionary and size
|
131 |
# each token directly reads off the logits for the next token from a lookup table
|
132 |
# see more: https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html
|
|
|
133 |
self.vocab_size = kwargs.get("vocab_size", 100)
|
134 |
self.num_embed = kwargs.get("num_embed", 32)
|
|
|
135 |
self.num_heads = kwargs.get("num_heads", 4)
|
|
|
136 |
self.num_layers = kwargs.get("num_layers", 4)
|
137 |
+
self.max_seq_len = kwargs.get("max_seq_len", 1024)
|
138 |
self.dropout = kwargs.get("dropout", 0.2)
|
|
|
139 |
# each token reads the logits for the next token from a lookup table
|
140 |
self.token_embedding_table = nn.Embedding(self.vocab_size, self.num_embed)
|
141 |
# each position from 0 to block_size-1 will get its embedding
|
142 |
+
#self.position_embedding_table = nn.Embedding(self.block_size, self.num_embed)
|
143 |
+
self.blocks = nn.ModuleList([
|
144 |
+
TransformerBlock(
|
145 |
+
num_heads=self.num_heads,
|
146 |
+
num_embed=self.num_embed,
|
147 |
+
dropout=self.dropout
|
148 |
+
)
|
149 |
+
for _ in range(self.num_layers)
|
150 |
+
])
|
151 |
+
# we add the layer norm before the Linear layer
|
|
|
|
|
|
|
|
|
152 |
self.lm_head = nn.Linear(self.num_embed, self.vocab_size)
|
153 |
+
self.norm = RMSNorm(self.num_embed)
|
154 |
+
|
155 |
+
self.freqs_cis = precompute_freqs_cis(
|
156 |
+
self.num_embed//self.num_heads,
|
157 |
+
self.max_seq_len * 2,
|
158 |
+
500000,
|
159 |
+
)
|
160 |
|
161 |
def forward(self, idx, targets=None):
|
162 |
B, T = idx.shape
|
163 |
# idx and targets are (B,T) tensor of integers
|
164 |
# the token_emb is (B, T, C), C = NUM_EMBED
|
165 |
x = self.token_embedding_table(idx)
|
|
|
|
|
|
|
|
|
166 |
|
167 |
+
freq = self.freqs_cis[:self.max_seq_len]
|
168 |
+
# apply one head of self-attention
|
169 |
+
for block in self.blocks:
|
170 |
+
x = block(x, freq)
|
171 |
|
172 |
+
x = self.norm(x)
|
173 |
+
|
174 |
# (B, T, vocab_size)
|
175 |
logits = self.lm_head(x)
|
176 |
+
# compute the loss
|
|
|
177 |
if targets != None:
|
178 |
# cross_entropy accepts inputs in a (batch_size, num_classes)
|
179 |
# so we need to reformat our logits dimensions to
|
180 |
# (batch_size * time, dim_vocabulary), time = block_size
|
|
|
|
|
181 |
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
|
182 |
else:
|
183 |
loss = None
|
|
|
184 |
return logits, loss
|
185 |
|
186 |
+
def generate(self, idx: torch.Tensor, max_new_tokens: int, temperature: float = 0.7, top_p: float = 0.9):
|
187 |
for _ in range(max_new_tokens):
|
188 |
idx_crop = idx[:, -self.max_seq_len:]
|
189 |
|
190 |
+
freq = self.freqs_cis[:self.max_seq_len]
|
191 |
logits, loss = self.forward(idx_crop)
|
192 |
logits = logits[:, -1, :]
|
193 |
|
194 |
if temperature > 0:
|
195 |
+
probs = F.softmax(logits / temperature, dim=-1)
|
196 |
idx_next = self.sample_top_p(probs, top_p)
|
197 |
else:
|
198 |
probs = F.softmax(logits, dim=-1)
|
199 |
idx_next = torch.multinomial(probs, num_samples=1)
|
200 |
idx = torch.cat((idx, idx_next), dim=1) # (B, T+1)
|
201 |
+
return idx[0]
|
202 |
|
203 |
def sample_top_p(self, probs: torch.Tensor, top_p: float) -> torch.Tensor:
|
204 |
sorted_probs, sorted_indices = torch.sort(probs, descending=True, dim=-1)
|
205 |
cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
|
206 |
+
|
207 |
# Create a mask for top-p filtering
|
208 |
top_p_mask = cumulative_probs <= top_p
|
209 |
top_p_mask[..., 1:] = top_p_mask[..., :-1].clone()
|